Driving Style Analysis Using an Improved Bidirectional Long Short-Term Memory Denoising Variational Autoencoder (I-Bi-LSTM-DVAE)

Traditional Long Short-Term Memory Autoencoders (LSTM-AE) often yield low-fidelity reconstructions, limiting the quality of learned representations for driving-style analysis, while Autoregressive Long Short-Term Memory (AR-LSTM) can suffer from information loss and error accumulation on long sequences. We propose an Improved Bidirectional Long Short-Term Memory Denoising Variational Autoencoder (I-Bi-LSTM-DVAE) that extracts driving characteristics from noisy trajectory data and enables accurate, data-driven driving-style recognition. The model employs a two-layer Bidirectional Long Short-Term Memory (Bi-LSTM) encoder within a Denoising Variational Autoencoder (DVAE) to learn continuous, noise-resistant latent distributions. Reconstruction is performed by a novel AR-LSTM decoder with a gating mechanism that fuses encoder features at every step, mitigating drift and preserving temporal consistency. Features learned by the encoder are clustered with K-means++ and evaluated on the NGSIM US-101 and I-80 datasets. Using standard criteria—Calinski–Harabasz (CH) index, Silhouette Coefficient, and Sum of Squared Errors (SSE)—the optimal number of clusters is consistently determined to be three, corresponding to aggressive, normal, and conservative styles. Our approach achieves a CH score of 28,115.66 and a silhouette score of 0.6752, outperforming statistical baselines, Principal Component Analysis (PCA), LSTM-AE, and denoising/autoregressive variants, demonstrating the method’s effectiveness for robust driving-style recognition and its potential to support safety analysis, intention recognition, and trajectory prediction.